Economics · 8 min read
Why agent token costs are the new cloud bill
A decade ago, the surprise line item on every engineering org's budget was cloud compute. Usage-based, easy to start, and — without guardrails — easy to let run away. Coding agents are doing the same thing to the model bill, only faster.
The anatomy of an agent invoice
A human developer sends a model a handful of prompts an hour. An agent sends hundreds. It reads files, plans, calls tools, re-reads its own output, retries on failure, and carries a growing context window through all of it. Each of those steps is billed in tokens — and the context is re-sent on nearly every call.
Three properties make this compound quickly:
- Context re-transmission. The same system prompt, tool definitions and file contents ride along on every step of a task.
- Tool output bloat. A single database query or file read can inject thousands of tokens the model barely needs.
- Flagship-by-default. Most agents point at one expensive model for everything, including work a far cheaper model would nail.
Why it grows with productivity
The uncomfortable part: the better your agents get, the more they run — and the higher the bill climbs. Spend scales with success, not waste. That's exactly the dynamic that made cloud costs a board-level topic, and it's why "just use fewer agents" is a non-answer.
The bill isn't a sign your agents are broken. It's a sign they're working — which is why you need a control plane, not a diet.
The three levers
You don't need to slow your agents down. You need to make each request cheaper without changing the outcome. There are three levers, and they compound:
- Compression. Fold repeated context and bloated tool output into a denser token stream, losslessly. Roughly half the input tokens, same answers. See token compression.
- Routing. Score each request and send simple work to an economy model, hard work to a flagship. Up to 15× cheaper on the requests that don't need the big model. See smart routing.
- Governance. Cache identical intent, cap budgets per team, and stop bad requests before they're billed.
Stacked together on every request, these routinely take 60–90% off the underlying bill — which is the whole reason Cohesor exists.